fix: keep hasToolResult record/replay symmetric + Anthropic tool_result+text scoping

[jpr5]

Problem

PR #316 (released as v1.37.3) scoped the replay-side hasToolResult check to the current turn (messages after the last user message) but left two paths inconsistent, so v1.37.3 can silently produce fixtures that never match their own recorded request.

Root cause

1. Record/replay split-brain. The matcher (src/router.ts) was changed to a turn-scoped check, but the recorder (src/recorder.ts, buildFixtureMatch) still STAMPED hasToolResult with a whole-conversation predicate messages.some(m => m.role === "tool"). Both operate on the same normalized ChatCompletionRequest.messages, so for a genuinely-recorded turn-2 leg-1 request [user, assistant, tool, user] (a fresh question whose history still carries turn-1's tool result):

• Recorder writes hasToolResult: true (whole conversation has a tool).
• Matcher on replay computes false (nothing after the last user) → true !== false → the fixture can never match its own recorded request. The comment at recorder.ts even claimed the value was "matcher-aware" — it no longer was.

2. Anthropic tool_result + text ordering. Anthropic bundles a leg-2 tool_result and accompanying user text into ONE user message. Normalization emitted [..., tool, user(text)], which is byte-identical to a genuine leg-1 turn, so the turn-scoped check saw no tool after the last user and misclassified a real leg-2 turn as false. Because the two shapes are indistinguishable after normalization, this can only be fixed where the source-message grouping is still known — in normalization, not in the shared helper.

Fix

• Extracted one shared currentTurnHasToolResult(messages) helper in src/router.ts (where the other message-shape utilities live and which recorder.ts already imports) and used it on both the record side and the match side. Sharing a single predicate makes the two sides structurally incapable of drifting. Replaced the false "matcher-aware" comment.
• Fixed claudeToCompletionRequest to emit the accompanying text user message before the tool message(s), so a real Anthropic leg-2-with-text turn normalizes to [..., user(text), tool] and is classified true, while a genuine leg-1 turn (separate messages) stays false. The common tool_result-only follow-up is unchanged.

Red-green proof

Both defects are proven on the real HTTP record→replay surface (spin up a mock upstream, record through the real recorder proxy to disk, replay the recorded fixtures through the real server/matcher). Source fixes were git stash-ed to observe RED, then restored for GREEN — tests unchanged between runs.

RED (source fixes reverted)

FAIL  hasToolResult record→replay symmetry > replays a genuinely-recorded turn-2 leg-1 request against its own recorded fixture
AssertionError: expected 503 to be 200 // Object.is equality
- Expected  200
+ Received  503

FAIL  hasToolResult record→replay symmetry > replays an Anthropic leg-2 turn that bundles tool_result WITH accompanying text
AssertionError: expected 503 to be 200 // Object.is equality
- Expected  200
+ Received  503

FAIL  claudeToCompletionRequest > handles tool_result with text blocks alongside in same user message
AssertionError: expected 'tool' to be 'user' // Object.is equality
Expected: "user"
Received: "tool"

(503 = strict-mode "No fixture matched": the recorded fixture could not match its own request.)

GREEN (source fixes restored, tests unchanged)

✓ src/__tests__/hastoolresult-record-replay-symmetry.test.ts (2 tests)
✓ src/__tests__/router.test.ts (144 tests)
✓ src/__tests__/messages.test.ts (73 tests)
✓ src/__tests__/thinking-invariants.test.ts (47 tests)
Test Files  4 passed (4)
     Tests  266 passed (266)

Both the OpenAI split-brain (Finding 1) and the Anthropic tool_result+text case (Finding 2) are covered by the real-path HTTP record→replay test, so both got a genuine RED→GREEN on the actual failure surface (not tests-against-fakes).

Tests added

• src/__tests__/hastoolresult-record-replay-symmetry.test.ts — full HTTP record→replay for (a) an OpenAI turn-2 leg-1 request and (b) an Anthropic tool_result+text leg-2 turn; both would 503 pre-fix.
• router.test.ts — direct currentTurnHasToolResult coverage (leg-2 true; new-turn turn-scoped false; leading-system-message; lastUserIdx === -1 no-user whole-conversation fallback), plus matcher composition of the scoping with explicit turnIndex and sequenceIndex gates, and a system-message-led new-turn case.
• messages.test.ts — the two existing tool_result+text tests that encoded the old (buggy) tool-before-text order were corrected to the fixed order.

Full suite

pnpm format:check, pnpm lint, pnpm build all clean. pnpm test: 4683 passed, 1 pre-existing wall-clock flake (timing-replay.test.ts > replaySpeed 2.0 halves the replay duration, expected 209 to be less than 200) that passes in isolation and is unrelated to this change.

Notes

No version bump and no changeset (repo has no .changeset/; releases are cut by a separate release PR). This completes the incomplete fix from #316.

🤖 Generated with Claude Code

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@319

commit: e192e5b